Chinese AI startup DeepSeek has just officially released its latest large language model (LLM), DeepSeek-V3-0324. With a capacity of up to 641GB, this model was announced on the Hugging Face AI platform without much information being revealed, in line with the company's usual secretive style about new products.
What makes this model unique is that it is licensed under the MIT license, which allows it to be used freely for commercial purposes. Initial benchmark results show that DeepSeek-V3-0324 is capable of running on mainstream hardware configurations, such as Apple’s Mac Studio with an M3 Ultra chip. AI scientist Awni Hannun reported that it was possible to achieve a processing speed of more than 20 tokens per second using this configuration. This ability to run a large language model on commodity hardware on-premises is a stark contrast to the traditional approach of using massive data center infrastructure to support advanced AI models.
According to DeepSeek, initial tests have shown significant improvements over previous versions. The model has been rigorously tested by internal stakeholders and has demonstrated excellent performance, even outperforming all competing models and beating Anthropic’s Claude Sonnet 3.5 on tasks that do not require deep learning. However, unlike Sonnet-style models that require a paid subscription, DeepSeek-V3-0324 is completely free to download and use.
Technically, DeepSeek-V3-0324 uses a mixture of experts (MoE) architecture. It is notable for its selective use of approximately 37 billion of the 685 billion parameters for each task, which increases efficiency by reducing computational demands while maintaining performance. The model also applies Multi-Head Latent Attention (MLA) and Multi-Token Prediction (MTP) technologies, which contribute to improved context memory and faster output.
Users can access DeepSeek-V3-0324 via Hugging Face, OpenRouter's chat interface and API, and DeepSeek's chat platform if they wish. Inference service provider Hyperbolic Labs also provides access to the model.
With the release of DeepSeek-V3-0324, the company continues to assert its position in the race to develop large language models, while providing a powerful and accessible option for the AI research and development community. The ability to run on commodity hardware, coupled with a free-to-use license, will certainly open up many opportunities for practical applications for this model in the near future.
Compare realistic side hustles for 2026 by startup cost, fees, flexibility, and income control, with current tax and scam guidance from official sources.
Upgrade your home office ergonomics with current OSHA and NIOSH guidance on chair fit, monitor height, laptop setup, lighting, and movement.
Learn why emotional intelligence matters in leadership, where it helps most, how to strengthen it, and how to tell whether your approach is working.
A practical roadmap for moving into tech without a degree: choose an accessible role, build job-ready skills, prove them with projects, and use hiring feedback to improve.
Build a credible LinkedIn personal brand in 2026 with current profile, content, newsletter, analytics, verification, and visibility guidance from LinkedIn.
Quiet quitting and quiet firing are catchy labels, not diagnoses. Learn what is verified, what depends on context, and what employees and managers can do next.
Fractional executives offer senior leadership without a full-time hire. Learn when the model works, what to ask, and the risks to manage.
Prepare a stronger Q4 raise request with current wage data, timing options, compensation tradeoffs, and scripts for yes, partial, or no outcomes.
Choose a high-income skill you can realistically start before 2026 ends, with current pay data, learning priorities, and a practical self-check.
Build a hybrid work schedule around tasks, team overlap, and measurable outcomes. See sample patterns, warning signs, and when to adjust your plan.
